shuaishuaicdp/game2world-agent-eval

Game2World Agent Evaluation

Recorded agent evaluations of GameCleaner and gameplay video editing baselines, including evidence images, prompts, judge responses, scores, and model output videos.

The collection contains 15,407 episodes across 18 evaluation runs, with 933,310 evidence images. Of these, 15,360 episodes contain both a captured prompt and a complete judgment; the remaining 47 partial records are retained for inspection.

Contents

Each evaluation episode is a separate ZIP, grouped into approximately 4 GB TAR files under episodes/. episodes.jsonl identifies each episode's run, video, variant, archive, and available records. Historical run names are kept so results can be traced back to their experiments.

An episode includes its evidence images and metadata, case_score.json, and the original files under judge/ where available. prompt.txt contains the prompt captured in the log; episode.json records which fields are present. Failed and partial evaluations are retained and marked in the index.

The logs contain only what was actually recorded. They are not a reconstruction of hidden reasoning, and some evaluations have no recorded reasoning or tool calls.

reports.zip contains aggregate evaluation tables and summaries. model-outputs/ contains edited videos and inference records, grouped by run. hud-quality-evaluation.zip preserves the separate HUD dataset quality evaluations.

supplemental/hf-live-judge-test.zip preserves one additional single-case release test, outside the main episode index.

Download

The full archive is about 384 GB. To start with one episode shard:

pip install -U huggingface_hub
hf download shuaishuaicdp/game2world-agent-eval --repo-type dataset \
  --local-dir game2world-agent-eval \
  --include README.md unpack_episode.py episodes.jsonl episodes/part-0000.tar

Omit --include and the filenames to download the complete collection.

To unpack an individual episode from a downloaded archive:

python game2world-agent-eval/unpack_episode.py --data game2world-agent-eval --output episode

This unpacks the first episode. Use --episode EPISODE_ID to select another record from episodes.jsonl.

When the original command is available, the script also writes image_order.json with the judge's recorded input order. Do not substitute alphabetical filename order when preparing multimodal training examples.

For judge fine-tuning or distillation, start with records that have both complete_judgment and captured_prompt. Split by source video, not by episode: the same video appears in multiple methods and evaluation runs.

Privacy

Text records are screened for credentials and personal paths before publication; local home directories are replaced with /workspace. Image metadata is inspected, and gameplay evidence is visually spot-checked. This is not a frame-by-frame privacy annotation of the underlying gameplay footage.

Related resources

• Game2World code
• Game2World benchmark
• GameCleaner LoRAs